Refresh invalidated Codex MCP clients without reloading configuration

[SaladDay]

Cancelling an active stdio MCP call reloads global and per-thread configuration after stopping its remote process scope. Those filesystem reads can exhaust the cancellation budget even after interruption and termination succeed. Mark the captured native Session’s MCP runtime dirty during targeted invalidation, then let existing reconciliation replace closed clients before subsequent use. Remove the adapter reload request and obsolete reload-failure fixture; retain exact scope closure, unaffected clients and the owned thread/workspace.

Validation:

• Complete Codex Go package tests, focused cancellation tests, vet and pinned-source patch applicability pass. Fresh independent review found no substantiated issues; all applicable CI checks pass.
• Official native build passes existing targeted-invalidation/binding Rust tests, initialize capability verification and real upstream V8 stdio execution.
• Real native/View/File/Process/Network fixture with three stdio services reproduces the original 10-second deadline in reload: 53 File metadata operations with 250 ms cancellation-only service delay. Simply removing reload makes the next target call fail, while healthy calls succeed.
• The complete candidate cancels in 30.285 ms without injected delay and 20.709 ms with the same 250 ms cancellation-only delay, with no File operations in cancellation. A subsequent Turn through the same Executor completes all three original calls: selected PID/start-time changes, both healthy identities stay unchanged, the old writer remains absent, its output stays stable and workspace outputs remain.

The controlled local results establish the mechanism and fix, not the sole cause of the recorded micro timeout or cloud qualification. Live acceptance remains part of final distribution qualification.